//! AMPL `.nl` reader, reverse-mode AD tape, and TNLP evaluator.
//!
//! This crate holds the `.nl` pipeline that used to live in `pounce-cli`:
//!
//! - [`nl_reader`] parses a `.nl` file into an [`nl_reader::NlProblem`] (the
//! `Expr` DAG, linear parts, bounds, names, starting point) and provides
//! [`nl_reader::NlTnlp`], a [`pounce_nlp::tnlp::TNLP`] implementation that
//! evaluates objective/gradient/Hessian and constraints/Jacobian.
//! - [`nl_tape`] flattens an `Expr` DAG into a reverse-mode AD tape with
//! colored forward-over-reverse Hessian products.
//! - [`nl_external`] supports AMPL imported (external) functions via the
//! `funcadd_ASL` ABI.
//! - [`nl_fbbt_translate`] lowers an `Expr` to an `FbbtTape` for
//! feasibility-based bound tightening.
//! - [`nl_quadratic`] recognizes a degree-≤2 polynomial in an `Expr` DAG and
//! reads off its Hessian, linear part, and constant — the algebra behind
//! the CLI's LP/QP/QCQP routing, kept here so the evaluator and the
//! extractors can share it.
//! - [`sol_writer`] is the reader's inverse: it formats a solve's primals,
//! duals, and suffixes as an AMPL `.sol` file. It lives here (rather than
//! in the CLI, which owned it until the browser frontend needed it too) so
//! every frontend emits the same file — same dual sign convention, same
//! suffix headers.
//!
//! It is a leaf crate (depends only on `pounce-common` and `pounce-nlp`) so
//! both the CLI and the Python bindings can read and evaluate `.nl` models
//! without depending on each other.